Research map / Image generation / Autoregressive and latent generation
Geometric Representation Condition Improves Equivariant Molecule Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GeoRCG, a two-stage molecule generator that first samples a pretrained geometric-encoder representation with a lightweight MLP diffusion model and then generates the 3D molecule conditioned on it, improving EDM and SemlaFlow on QM9/GEOM-DRUG and conditional generation.
Problem
Equivariant 3D molecule generative models struggle to produce high-quality molecules, especially under conditional property constraints, because the extrinsic molecular distribution is symmetric and high-dimensional.
Approach
Decompose molecular generation into a representation stage (diffusion over a symmetry-free embedding from a frozen pretrained geometric encoder, conditioned on atom count) and a molecule stage (equivariant diffusion conditioned on that representation), enabling parallel training and property-conditioned generation by retraining only the representation generator.
Key results
- On QM9 unconditional generation, GeoRCG (EDM) raises molecule stability from 82% to 92.32% and validity from 91.9% to 96.52%.
- On conditional QM9 generation, GeoRCG (EDM) reduces MSE error by an average of about 50% relative to the best prior models (e.g., alpha error 0.86 vs 1.97 for GCDM).
- With 100 diffusion steps, GeoRCG (EDM) reaches 91.85% molecule stability on QM9, surpassing EDM and GeoBFN at all step counts tested.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add geometric-representation conditioning to EDM | EDM · QM9 unconditional generation | Mol Sta (%): 82 -> 92.32(0.06) | Helped |
| Add geometric-representation conditioning to EDM | EDM · GEOM-DRUG unconditional generation | Atom Sta (%): 81.3 -> 84.3(0.12) | Helped |
| Add geometric-representation conditioning to SemlaFlow | SemlaFlow · GEOM-DRUG unconditional generation | Valid (%): 94.4(0.17) -> 95.3(0.13) | Helped |
| Add geometric-representation conditioning to EDM with target-property representations | EDM · QM9 conditional generation | MSE: 2.76 -> 0.86(0.01) (α); 655 -> 325.2(3.4) (Δε); 356 -> 202.2(1.2) (εHOMO); 584 -> 257.9(5.5) (εLUMO); 1.111 -> 0.805(0.0 | Helped |
| Use geometric representation conditioning with 100 diffusion steps | EDM · QM9 unconditional generation | Mol Sta (%): 69.8(0.2) -> 91.85(0.34) | Helped |
| Perturb the conditioning representation with Gaussian noise during molecule-generator training (ablation) | GeoRCG molecule generator · QM9 unconditional generation using sampled representations | Mol Sta (%) | Helped |
| Increase classifier-free guidance strength and decrease representation temperature (ablation) | GeoRCG (EDM) · QM9 unconditional generation | stability and uniqueness: 93.9%; 86.82% | Mixed |
Surprising findings
- Training the molecule generator on exact encoder representations overfits: sampled representations drop molecule stability from 93.2% to 86.9% without representation perturbation.
Limitations
- Generation quality depends heavily on encoder representation quality; weak pretraining (e.g., on GEOM-DRUG) limits gains from low-temperature sampling and CFG.
- Classifier-free guidance doubles batch size, roughly doubling memory and compute.
- Conditional-generation novelty concern: the representation generator may reproduce training molecules, though reported novelty is comparable to other methods.
Experimental setup
Datasets: QM9 GEOM-DRUG PCQM4Mv2
Benchmarks: QM9 GEOM-DRUG
Metrics: atom stability molecule stability validity validity and uniqueness energy strain energy MAE property error BondLengthW1 BondAngleW1
Baselines: EDM EDM-Bridge GeoLDM GCDM EquiFM GOAT GeoBFN SemlaFlow MiDi EQGAT-diff
Builds on: EDM RCG SemlaFlow Frad Chroma
Closest papers by meaning
- Learning Disentangled Equivariant Representation for Explicitly Controllable 3D Molecule Generation (AAAI 2025 · similarity 0.89)
- Unified Generative Modeling of 3D Molecules with Bayesian Flow Networks (ICLR 2024 · similarity 0.89)
- Enhancing Generalizability in Molecular Conformation Generation with METRIZATION-Informed Geometric Diffusion Pretraining (AAAI 2025 · similarity 0.88)
- Geometry Informed Tokenization of Molecules for Language Model Generation (ICML 2025 · similarity 0.88)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.88)
- Unified Guidance for Geometry-Conditioned Molecular Generation (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.